MOSAIC: Query-Aware Exploration Policy Adaptation for GraphRAG
EunKyeong Lee, Kyeong-Jin Oh, Jinwon Kim, Hye Woo Lee, Minsang Song, Hyeongjun Jang, Junyoung Youn
cs.AI
Quick verdict
GraphRAG as per-query policy control: strong controlled gains on GraphRAG-Bench, but externally reported baselines and proprietary code limit verification.
Summary
Mosaic (KT Corporation) is a training-free GraphRAG framework that treats retrieval as a per-query control problem. An LLM analyzer (GPT-4o-mini) maps each query to a bounded policy over seed selection, traversal, cumulative-gain stopping, and evidence selection, while the corpus graph, indexes, and generator stay fixed; six extensible signals (e.g., mediatedness, output cardinality) drive the analyzer, with allowlist validation and clamping. On GraphRAG-Bench, Mosaic reaches query-weighted Answer Correctness of 76.97 (Medical) and 64.33 (Novel), exceeding the strongest previously reported results by 5.13 and 4.43 points, with 95.1 Evidence Recall on Medical. Controlled comparisons on an identical graph and generator show no fixed narrow/medium/wide policy is uniformly best; Mosaic beats the strongest fixed policy by 9.96 points while evaluating 81.9% fewer paths than Fixed Wide, though end-to-end latency rises from 5.29s to 9.68s. The policy interface transfers to HotpotQA, MuSiQue, and 2WikiMultiHopQA without benchmark-specific training, but a −12.6 EM gap remains on 2WikiMultiHopQA. Main limitations: baselines are externally reported rather than rerun, the implementation is proprietary with only promised controlled-access artifacts, the analyzer can misclassify evidence requirements (a documented failure case), and no statistical significance testing is reported.
Review complete
Assessment at a glance
6
risk flags identified
● high 0
● medium 4
● low 2
4 strengths · 4 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Query-specific policy construction beats every fixed operating point on a shared graph and generator, by 9.96 points over the strongest canonical fixed policy on Medical (Table 6).
- On GraphRAG-Bench, Mosaic achieves query-weighted Answer Correctness of 76.97 (Medical) and 64.33 (Novel), exceeding the strongest previously reported overall results by 5.13 and 4.43 points (Table 4).
- The gains come from better allocation rather than more search: 81.9% fewer evaluated paths and 47.2% fewer retained evidence items than Fixed Wide (Table 9).
- The bounded policy interface transfers to HotpotQA, MuSiQue, and 2WikiMultiHopQA without benchmark-specific retriever training, though not at SOTA level (§8).
Strengths
Controlled fixed-policy ablation isolates the contribution
Table 6 runs Narrow/Medium/Wide/Budget-Matched policies on the identical graph, generator, and retrieval code; the Budget-Matched control (72.03) is especially informative, since Mosaic's remaining +4.94 over it approximates the true policy-construction effect rather than budget effects.
Unusually candid self-assessment
§10.2 enumerates what the results do not establish (no per-phrase causality, non-matched baseline reproductions, asymmetric transfer conditions); Table 11 includes a genuine failure case (SLNB: Mosaic 24% vs Narrow 99%); Table 8 admits latency nearly doubles (9.68s vs 5.29s).
Policy diversity and alignment analysis goes beyond averages
Efficiency evidence rules out brute-force explanations
Weaknesses
Headline SOTA deltas rest on externally reported baselines
Reproducibility is currently nominal
No statistical significance testing for any accuracy delta
Single analyzer model and benchmark-derived calibration
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
Mosaic (KT Corporation, arXiv 2609.11065) is a technical report proposing a training-free GraphRAG framework that reframes retrieval as a per-query control problem. An LLM analyzer (GPT-4o-mini) converts a query's evidence requirements into a bounded, validated policy πq covering seed selection, traversal (depth/width/anchoring/connectors), cumulative-gain stopping, and evidence retention (§5, Table 3). Six composable signals — Mediatedness, Output Cardinality, Seed Coverage, Anchoring Need, Identity–Relation Dependence, Completeness — drive the analyzer, which the authors explicitly frame as an extensible interface, not a closed taxonomy (§4). Headline claims: (i) query-weighted Answer Correctness of 76.97 on Medical and 64.33 on Novel, +5.13/+4.43 over the strongest previously reported systems (Table 4); (ii) controlled evidence that no fixed narrow/medium/wide policy is uniformly optimal, with Mosaic +9.96 over the strongest fixed policy on an identical graph and generator (Table 6); (iii) efficiency — 81.9% fewer evaluated paths and 47.2% fewer retained evidence items than Fixed Wide (Table 9); (iv) training-free transfer to three multi-hop QA benchmarks (Table 10).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 4 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is MOSAIC?
MOSAIC is a training-free framework for Graph Retrieval-Augmented Generation (GraphRAG) from KT Corporation. Instead of using one fixed retrieval configuration for all queries, an LLM analyzer reads each question and emits a bounded 'policy' controlling seed selection, graph traversal depth/width, when to stop exploring, and how much evidence to keep. The underlying corpus graph, indexes, and answer generator stay unchanged.
How much better is MOSAIC than existing GraphRAG systems?
On the GraphRAG-Bench benchmark, MOSAIC reports query-weighted Answer Correctness of 76.97 on the Medical split and 64.33 on the Novel split — 5.13 and 4.43 points above the strongest previously published results (G-Reasoner). In the authors' own controlled comparison on an identical graph and generator, it beats the best fixed retrieval policy by 9.96 points. Note these baseline numbers were taken from prior publications, not rerun by the authors.
Does MOSAIC make GraphRAG faster or cheaper?
No. End-to-end latency rises from 5.29 seconds (fixed medium policy) to 9.68 seconds because the analyzer adds a 2.82-second LLM call per query. The efficiency claim is narrower: MOSAIC explores 81.9% fewer graph paths and retains 47.2% less evidence than a uniformly wide policy while scoring higher, so it uses less graph search per unit of accuracy.
Is MOSAIC's code available?
Not currently. The paper states the implementation contains proprietary components and that the authors intend to release benchmark configurations, retrieval code, and outputs through a controlled-access repository subject to organizational approval. The GraphRAG-Bench benchmark itself is public, and the paper specifies most hyperparameters, but the analyzer prompts are not included in the text.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.